“Google's "Chinchilla" paper established an optimal ratio of data to model parameters for training large language models.”